Figure 4 shows the latency of three different size Clos networks as a function of the aggregate network throughput. The traffic pattern is random, i.e. transmitting nodes choose a destination from a uniform distribution. The packet length is 64 bytes. The results are produced by varying the network load and measuring the corresponding throughput and latency. It can be seen that the average latency increases exponentially as the network throughput approaches saturation. Therefore, to achieve low average latencies the network load must be below the saturation throughput.
Figure 5 shows the probability that a packet will have a latency greater
than a given value for various network loads. The traffic pattern
is random, with a packet length of 64 bytes. For 10% load the latency
distribution is very narrow compared to higher loads. Near the saturation
throughput (about 60% load) a significant percentage of the packets
experience a latency many times the average value, which is 18 
s.
To reduce the probability of very large latency values the network load must be
far below the saturation throughput.
Figure 5: Probability that a packet will have greater than a given
latency value for a 64 node Clos network
Figure 4: Latency versus throughput for 64, 128 and 256 node Clos networks